目标函数对神经网络性能影响的研究

通过对常用目标函数的性能，网络的收敛性及其统计特性影响的系统分析，给出了基于李雅普洛夫定理的有关收敛性的理论证明。理论分析和相关实验表明，神经网络采用选择性强的目标函数其逼近能力也强，而选择合适的目标函数，可以保证神经网络逼近是对函数的无偏估计。

Effect of Object Function on the Performance of the Neural Network

The capability of the general object functions is introduced. The effect of the object function on the statistical performance and astringency of the neural network (NN) is analyzed. Based on Lyapinov theorem, the astringency of the NN is proved under the general object function. The approximate speed and the statistical performance are explored under the different object function. The theoretica analysis and experiments indicate that the astringency can be improved through the better selectivity object function, and the appropriate object function can guarantee that NN approximates to the function completely.